AI GPU Cloud Comparison 2026
Compare RunPod, Vast.ai, Lambda and Hetzner for GPU rentals. Start with GPU memory, billing terms and your workload—not a cheapest-provider badge.
Four GPU rental options
TL;DR: Match the exact GPU and memory first. Hourly instances, marketplace offers and dedicated servers are different products. These are documented options, not a performance ranking. Official sources checked 13 October 2026; availability and checkout prices can change.
GPU instances
RunPod
$2.89/GPU-hour (H100 PCIe 80GB)
- Pods list: H100 PCIe 80GB $2.89/hour
- RTX 4090 24GB listed at $0.89/hour
- Confirm cloud tier, region, storage and separate Serverless rates
Vast.ai
Vast.ai
Live quotes / marketplace
- Hosts set prices; no universal H100 rate
- Compare dedicated versus interruptible offers
- Budget for compute, storage and bandwidth
Lambda
Lambda
$3.29/GPU-hour (H100 PCIe 80GB)
- 1-GPU H100 PCIe 80GB: $3.29/GPU-hour
- 1-GPU H100 SXM 80GB: $4.29/GPU-hour
- Minute billing; applicable taxes extra; capacity not guaranteed
Dedicated GPU servers
Hetzner GPU
Check configuration
- GEX45: RTX PRO 4000 Blackwell SFF
- GEX63/GEX131: 96GB VRAM per server
- GPU locations: Germany or Finland, depending on model
Frequently Asked Questions
How do I compare H100 quotes fairly?
Compare PCIe with PCIe or SXM with SXM, then match GPU count, region and rental type. A per-GPU price for an eight-GPU instance is not the total instance bill. Add storage, transfers and taxes where applicable. This page does not establish a market-wide lowest price or measure provider reliability.
Do I need H100 or A100 for fine-tuning?
There is no universal requirement based on parameter count alone. Quantization, full versus adapter training, sequence length, batch size and optimizer state affect memory. First reproduce your intended configuration on a short rental and record peak memory and runtime. Hugging Face explains quantization with PEFT; its examples are not a guarantee that every model fits your GPU.
Serverless vs on-demand GPU?
Compare idle-worker charges, minimum billing, cold starts and persistent storage—not just the label. Serverless can suit intermittent requests when its scaling policy matches your latency needs. A rented instance offers more direct control for sustained jobs. Hourly price displays do not necessarily mean hourly rounding: Lambda documents minute billing and Vast.ai documents second-based usage.
When should I consider a dedicated GPU server?
Consider one when your measured workload fits the offered GPU and you need sustained use. Hetzner lists single-GPU RTX-based GEX servers, not H100 instances. Confirm setup fees, monthly caps, location and tax in the selected configuration. Compare total monthly cost against actual rental hours; do not treat different GPUs as equivalent performance.